--- title: The open-source AI model landscape — June 2026 canonical: "https://agenticup.dev/posts/open-source-ai-model-landscape-june-2026/" pubDate: "2026-06-13T00:00:00.000Z" description: "Benchmarks tell you what's technically capable. Production adoption tells you what actually works at scale. These are two different lists. Here's the 2026 open-source landscape ranked for real workloads." tags: [open-source-llms, deepseek, kimi, qwen, glm, minimax, xiaomi, model-comparison, production-ai] --- TL;DR: The open-source model landscape in June 2026 has three distinct tiers — cheap inference models under $0.10/M tokens, capable mid-range models under $0.50/M, and frontier models that compete with Claude Opus 4.7 and GPT-5.5. Every model on this list comes from a Chinese AI lab. The cost of intelligence has collapsed by 10-100x in 12 months. > **Key takeaways:** > - DeepSeek V4-Flash and Xiaomi Mimo V2.5 lead the cheap inference tier at $0.09-0.14/M tokens > - Kimi K2.6 and DeepSeek V4-Pro dominate the mid-range for agentic coding > - Qwen 3.7 Max matches Claude Opus 4.7 on agentic benchmarks at half the price > - Every model here is open-weight — you can download, self-host, or fine-tune > - Chinese labs (8 of them) have released more open-weight models than the rest of the world combined in 2026 Benchmark scores tell you what's technically capable. Production adoption tells you what actually works at scale. These are two different lists. The AA Coding Index ranks DeepSeek V4-Pro at 47.5 and Kimi K2.6 at 47.1. But if you look at what's actually being called in production — API volumes, token consumption, provider infrastructure — the picture is different. The models winning real workloads are the ones that balance capability with cost, latency, and reliability. Here's the 2026 open-source model landscape, ranked by production utility. Three tiers. Nine models. ## The production tier (under $0.15/M tokens) ### 1. DeepSeek V4-Flash The most-called open-source model in production today. DeepSeek launched two V4 variants on April 23, 2026. V4-Flash is the smaller one — 284B total parameters, 13B active per token, packed into a 158GB weight footprint. It shares the same 1M-token context, MIT license, and core architecture as the Pro variant. **What makes it dominant:** Price. At $0.14/M input tokens and $0.28/M output, it's roughly 25x cheaper than GPT-5.5. Prompt caching drops that further — cache hits cost a fraction of cache misses, and in practice, most production workloads hit cache rates above 90%. The effective cost is near zero for predictable patterns. **What it handles:** - High-volume chat and completion workloads - Structured code generation from specs - Tool-calling and function-calling patterns - RAG pipelines with long-context retrieval - Batch processing and data extraction **Where it falls short:** It's not a frontier reasoning model. Complex multi-step agentic workflows, deep debugging, and tasks requiring sustained chain-of-thought still benefit from the Pro variant. You trade depth for speed and cost. **Specs:** 284B total / 13B active (MoE), 1M context, MIT license, text-only. Available via DeepSeek API. --- ### 2. Xiaomi Mimo V2.5 This was the surprise of 2026. Xiaomi, known for phones and IoT, released the Mimo V2 series on March 18, 2026 — and it's competitive. The V2.5 sits at the budget end: 262K context window, up to 65K output tokens per request. At $0.09/M input and $0.29/M output, it's the cheapest capable model on this list. Cache read drops to $0.04/M. **What it handles:** - Cost-sensitive production pipelines - Tool-calling and multi-step reasoning (support is native) - Code generation and review - Lightweight agent workflows **Where it falls short:** Context is 262K vs the 1M standard that's emerging. Xiaomi's API infrastructure is newer and less proven than DeepSeek's or Alibaba's. The model is open-weight but the broader ecosystem (documentation, SDKs, community) is thinner. **Specs:** 262K context, 65K max output, tool calling + reasoning support. Available via Xiaomi MiMo platform and hosted providers. --- ### 3. MiniMax M2.7 MiniMax released M2.7 as an agentic productivity model — designed for coding tools, multi-step office tasks, and agent harnesses rather than generic chat. It's positioned differently from the M3: M2.7 is the workaday variant with 200K context at $0.30/M input. That's still cheap, but notably more than the flash-tier models. What you get in return is better agentic reliability — the model was trained with tool-calling loops in mind. **What it handles:** - Code generation inside agent loops - Multi-round document editing - Integration with coding tools (Claude Code, Cursor, OpenClaw) - Structured output and function calling **Where it falls short:** 200K context is less than most competitors at this price point. The highspeed variant doubles the cost ($0.60/M). If you don't need agentic reliability, the flash models are cheaper. **Specs:** 200K context, 128K max output, $0.30/M input ($0.60/M highspeed), tool calling + caching. Available via MiniMax API and third-party providers. ## The capable tier ($0.15-$0.50/M tokens) ### 4. DeepSeek V4-Pro If V4-Flash is the production workhorse, V4-Pro is the capability upgrade. It routes 1.6 trillion total parameters through a much deeper expert pool — 49B active per token. The architecture is the same (MoE, Compressed Sparse Attention, 1M context), but the model has more specialized expert sub-networks for reasoning-heavy tasks. **The pricing story is worth understanding:** DeepSeek launched V4-Pro at a 75% discount — $0.435/M input, $0.87/M output — running through May 31, 2026. The regular rate ($1.74/M input, $3.48/M output) is higher, but still competitive with Western models. As of June, the promo has ended but many providers negotiated ongoing volume discounts. **How it compares to Flash:** On benchmarks, V4-Pro consistently scores 15-25% higher on coding, reasoning, and math tasks. In production, the gap depends on your workload. For straightforward completions and structured code, Flash handles 90% of what Pro does at 10% of the cost. For complex debugging, long-horizon planning, and agentic workflows, Pro earns its price. **Specs:** 1.6T total / 49B active (MoE), 1M context, MIT license, text-only (preview). --- ### 5. Kimi K2.6 Moonshot AI's Kimi K2.6 is the strongest agentic coder in the open-weight category. The AA Coding Index puts it at 47.1 — a nose behind DeepSeek V4-Pro's 47.5. But on Agentic SWE benchmarks (software engineering with multiple turns), K2.6 ties or beats V4-Pro. Moonshot trained it with multiple agent scaffolds to avoid overfitting to a single harness, and it shows in production. **What makes it different:** K2.6 was designed from the ground up as an agent model, not a chatbot that can also call tools. Its architecture prioritizes multi-turn consistency, tool selection accuracy, and error recovery — the failure modes that matter in agent loops. **The tradeoff:** It needs more compute per token than Flash models, and the 256K context is smaller than the 1M class. But if you're building agents, not chatbots, K2.6 is the strongest open-weight option. **Specs:** 256K context, strong agentic coding scores, Moonshot API. Open-weight. --- ### 6. Qwen 3.6 Plus Alibaba's Qwen 3.6 Plus is a closed-weight model available through Fireworks AI outside China. At $0.50/M input ($0.10/M cached) and $3.00/M output, it sits in the middle of the pricing spectrum. **The interesting angle:** It's a MoE model that supports vision — one of the few models at this price point that handles images natively alongside text. If you need multimodal without jumping to the frontier tier, this is the option. **What it handles:** - Text + image inputs in a single pipeline - Function calling and structured output - Production API workloads via Fireworks **Where it falls short:** It's closed-weight, so you can't download or self-host it. The ecosystem depends on Alibaba's infrastructure or Fireworks routing. For pure text workloads, DeepSeek and Kimi offer similar quality at lower prices. **Specs:** MoE architecture, context length unlisted, vision support, function calling. Available via Fireworks AI. Closed-weight. --- ### 7. MiniMax M3 M3 is MiniMax's answer to the question: what if you didn't have to choose between coding performance, long context, and multimodality? Released June 1, 2026, M3 scores 59.0% on SWE-Bench Pro (ahead of GPT-5.5's 58.6%), supports 1M tokens via MiniMax Sparse Attention, and handles text, images, and video natively. At $0.60/M input, it's priced between the capable and frontier tiers. **The sparse attention matters:** Full attention at 1M tokens is computationally prohibitive. M3's MSA architecture selects the most relevant KV blocks instead of attending to all of them, making long-context inference economically viable for the first time in an open-weights model. **Where it fits:** If you need 1M context and multimodal in a single model, M3 is the only open-weights option that delivers both. For pure text coding, Kimi K2.6 and DeepSeek V4-Pro score higher on agentic benchmarks. **Specs:** 1M context, 128K max output, $0.60/M input ($2.40/M output), modified-MIT license, text + image + video. ## The frontier tier ($0.50+/M tokens) ### 8. Qwen 3.7 Max Alibaba's Qwen 3.7 Max, released May 20, 2026, is the strongest Chinese challenger to Claude Opus 4.7 and GPT-5.5. The numbers are striking: SWE-Bench Pro 60.6 (beats Opus 4.7's 60.3), Terminal-Bench 2.0 at 69.7, GPQA at 92.4. On agentic coding benchmarks, it matches or beats every proprietary model. The AA Intelligence Index scores it at 56.6 Chinese — which means it was trained on Chinese-heavy data and may perform differently on English-heavy tasks. **Pricing:** $2.50/M input, $7.50/M output. That's half of Opus 4.7 on input, slightly more on output. Cached input drops to $0.25/M (90% off). OpenRouter shows $1.25/$3.75 as a launch promo — if that's still active, it's the best value on the frontier. **The catch:** It's text-only and closed-weight. No images, no self-hosting, no fine-tuning. It's also designed for extended-thinking mode — it uses inference-time compute to reason before responding, which adds latency. **When to use it:** Complex agentic coding tasks, multi-file refactors, debugging sessions that need sustained reasoning. Not for high-volume API calls — that's what Flash models are for. **Specs:** 1M context, 66K max output, $2.50/M input ($7.50/M output), extended-thinking mode, text-only, closed-weight. --- ### 9. Qwen 3.7 Plus Qwen 3.7 Plus is the affordable sibling. It costs a sixth of Max on input and handles images — the one capability Max doesn't have. At roughly $0.40-$0.50/M input (pricing varies by provider), it's the mid-range option in the Qwen 3.7 family. It shares the same architecture family but routes through fewer experts. If you need multimodal but can't justify Max's price, this is the Alibaba option. **Specs:** Vision support, function calling, available via Fireworks AI and DashScope. Closed-weight. ## The decision framework Picking the right model is about understanding what tier your workload falls into: | Workload Type | Recommended Model | Why | |---|---|---| | High-volume chat, batch processing, data extraction | DeepSeek V4-Flash or Mimo V2.5 | Lowest cost with prompt caching | | Agentic coding with tool calls | Kimi K2.6 or DeepSeek V4-Pro | Best agentic training, multi-turn reliability | | Single-agent coding assistant | MiniMax M2.7 | Purpose-built for agent harnesses | | Long-context + multimodal | MiniMax M3 | Only open-weights model with both | | Frontier reasoning + coding | Qwen 3.7 Max or DeepSeek V4-Pro | Beats proprietary models on agentic benchmarks | | Vision + text workflow | Qwen 3.6 Plus or Qwen 3.7 Plus | Multimodal at mid-range pricing | | Cost-sensitive production (>1M calls/month) | V4-Flash or Mimo V2.5 | Under $0.15/M with caching | ## What changed in 2026 Three things shifted the landscape this year. **First, the pricing collapse.** DeepSeek's V4 launch reset expectations. When a 284B-parameter model with 1M context costs $0.14/M, every other provider has to justify their pricing. The result is a market where capable inference costs less than $0.10/M tokens. **Second, the agentic training focus.** Kimi K2.6, MiniMax M2.7, and M3 were all trained with agent workflows in mind — multi-turn tool use, error recovery, structured output. The models that win in production are the ones that work reliably inside loops, not the ones that score highest on static benchmarks. **Third, the Chinese lab offensive.** Every model on this list comes from a Chinese AI lab. DeepSeek, Moonshot, Zhipu, Alibaba, MiniMax, and Xiaomi have collectively released more open-weight models in 2026 than all Western labs combined. The geopolitical implications aside, this means developers have more choice and lower prices than ever. The gap between open-source and proprietary has narrowed to 5-10% on structured tasks. For the workloads that make up 80% of production AI — code generation, RAG, classification, extraction — open-source models are now the default choice. The remaining gap is in long-context reasoning, creative work, and tasks that need the very best model regardless of cost. For those, Claude and GPT still lead. But the margin is shrinking every quarter. ## FAQ > **What is the best open-source AI model for production in June 2026?** > There's no single winner. DeepSeek V4-Flash dominates high-volume production because it costs $0.14/M tokens with massive prompt caching. For complex agentic work, Kimi K2.6 and DeepSeek V4-Pro lead. For the frontier tasks, Qwen 3.7 Max matches or beats Claude Opus 4.7 on agentic coding benchmarks at half the price. > **Which open-source model is cheapest?** > DeepSeek V4-Flash at $0.14/M input and Xiaomi Mimo V2.5 at $0.09/M input are the cheapest capable models available. Both support tool calling and reasoning. With prompt caching, effective costs drop by 90% or more. > **How do Chinese AI labs dominate open-source models?** > Every model on this list comes from a Chinese AI lab — DeepSeek, Moonshot, Zhipu, Alibaba (Qwen), MiniMax, and Xiaomi. Chinese labs have released more open-weight models than the rest of the world combined, and they compete aggressively on price. DeepSeek's V4 launch alone reset market pricing by 5-10x. > **Should I use an open-source model or a proprietary one?** > Open-source models now match or beat proprietary models for structured coding tasks, tool use, and high-volume API calls. Proprietary models (Claude, GPT) still lead on long-context reasoning, creative writing, and complex agentic workflows. The gap has narrowed to 5-10% on most benchmarks. ## Related Posts - [Best Open-Source LLMs for Coding 2026](/posts/best-open-source-llms-coding-2026/) — Benchmarks and technical comparison of top coding models - [MiniMax M3: open-weights coding, 1M context, and multimodality at 12x less than GPT](/posts/minimax-m3-review/) — Deep dive on MiniMax's latest model - [Best AI Coding Agents 2026: Ranked for Real Projects](/posts/best-ai-coding-agents-2026/) — The agent tools that run these models - [The Vertical Agent Method](/posts/the-vertical-agent-method-framework/) — How to pick the right workflow before you pick the model - [When a "worse" model beats a frontier model for agent work](/posts/when-better-model-isnt-better-agent/) — Why production economics matter more than benchmarks --- This article was published on Agentic Up (https://agenticup.dev) — practical guides for developers and founders building with AI agents. Reach me at hello@agenticup.dev.